Evaluation and Hydrological Application of CMADS Reanalysis Precipitation Data against Four Satellite Precipitation Products in the Upper Huaihe River Basin, China

Satellite- and reanalysis-based precipitation products are important data source for precipitation, particularly in areas with a sparse gauge network. Here, five open-access precipitation products, including the newly released China Meteorological Assimilation Driving Datasets for the Soil and Water Assessment Tool (SWAT) model (CMADS) reanalysis dataset and four widely used bias-adjusted satellite precipitation products [SPPs; i.e., Tropical Rainfall Measuring Mission (TRMM) Multisatellite Precipitation Analysis 3B42 Version 7 (TMPA 3B42V7), Climate Prediction Center (CPC) morphing technique satellite-gauge blended product (CMORPH-BLD), Climate Hazards Group Infrared Precipitation with Station Data (CHIRPS), and Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks-Climate Data Record (PERSIANN-CDR)], were assessed. These products were first compared with the gauge observed data collected for the upper Huaihe River basin, and then were used as forcing data for streamflow simulation by the Xin’anjiang (XAJ) hydrological model under two scenarios with different calibration procedures. The performance of CMADS precipitation product for the Chinese mainland was also assessed. The results show that: (1) for the statistical assessment, CMADS and CMORPH-BLD perform the best, followed by TMPA 3B42V7, CHIRPS, and PERSIANN-CDR, among which the correlation coefficient (CC) and root-mean-square error (RMSE) values of CMADS are optimal, although it exhibits certain significant negative relative bias (BIAS; −22.72%); (2) CMORPH-BLD performs the best in capturing and detecting rainfall events, while CMADS tends to underestimate heavy and torrential precipitation; (3) for streamflow simulation, the performance of using CMADS as input is very good, with the highest Nash-Sutcliffe efficiency (NSE) values (0.85 and 0.75 for calibration period and validation period, respectively); and (4) CMADS exhibits high accuracy in eastern China while with significant negative BIAS, and the performance declines from southeast to northwest. The statistical and hydrological evaluations show that CMADS and CMORPH-BLD have high potential for observing precipitation. As high negative BIAS values showed up in CMADS evaluation, further study on the error sources from original data and calibration algorithms is necessary. This study can serve as a reference for selecting precipitation products in data-scarce regions with similar climates and topography in the Global Precipitation Measurement (GPM) era.


Introduction
As a primary source of water resource on earth, precipitation plays a significant part in driving global en-ergy and hydrological cycles (Hou et al., 2014;Maggioni et al., 2016;Skofronick-Jackson et al., 2017). Accurate precipitation data are crucial for water resources management, flood forecasting, and drought monitoring (Jiang et al., 2012(Jiang et al., , 2014Wu et al., 2014;Sahoo et al., 2015;Guo et al., 2016;Maggioni and Massari, 2018;Zhong et al., 2019). Therefore, accurate and reliable globally covered precipitation records must be obtained in time. At present, there are three main ways to obtain precipitation datasets including observed gauge precipitation, satellite-related precipitation retrievals, and reanalysis precipitation datasets (Sun et al., 2018), in which rain gauge observations provide relative reliable precipitation data. However, as the sparseness and unevenness of rain gauges in some regions, and rain gauge precipitation exhibits high spatial heterogeneity, constant and accurate precipitation records cannot always be obtained. As such, satellite precipitation products (SPPs) and reanalysis precipitation datasets have been widely used owing to their high spatiotemporal resolution and wide cover range, and they are open access (Seyyedi et al., 2015;Duan et al., 2019;Zhu et al., 2019).
During the past 20 years, a number of SPPs, including the Tropical Rainfall Measuring Mission (TRMM) Multisatellite Precipitation Analysis (TMPA; Huffman et al., 2007), Climate Prediction Center (CPC) morphing technique product (CMORPH; Joyce et al., 2004), Climate Hazards Group Infrared Precipitation (CHIRP; Funk et al., 2015), and Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN; Sorooshian et al., 2000), have been operationally available. These SPPs were widely applied in hydrology and water resources research, and had great potential for streamflow simulation, flood frequency analysis, and extreme events monitoring at various regions, particularly for the bias-adjusted post-real-time products, including TMPA 3B42 Version 7 (TMPA 3B42V7), the CMORPH satellite-gauge blended product (CMORPH-BLD), CHIRP with Station Data (CHIRPS), and PER-SIANN-Climate Data Record (PERSIANN-CDR; Golian et al., 2015;Tan et al., 2015;Duan et al., 2016;Gao et al., 2017;Su et al., 2017;Jiang et al., 2018a;Wu et al., 2018;Zhang et al., 2018;Chen et al., 2019;Golian et al., 2019;Lai et al., 2019;Sulugodu and Deka, 2019;Zhang et al., 2019). Duan et al. (2016) evaluated eight high-spatial-resolution gridded precipitation products over Adige Basin in Italy, and concluded that the CHIRPS, TMPA 3B42V7, and CMORPH-BLD performed best. Su et al. (2017) assessed the applicability of four satellite-gauge combined precipitation products for predicting extreme precipitation and streamflow over the upper Yellow River basin, China, and pointed out that CMORPH-BLD agreed best with the gauge observations, and the simulated streamflow was comparable to the gauge-based result. Maggioni and Massari (2018) summarized previous studies using SPPs for riverine flood modeling and highlighted that SPPs have significant potential in flood forecasting, and precipitation bias correction and model recalibration were two viable options to improve SPPforced streamflow simulations.
As SPPs are estimated by calibrated infrared (IR), microwave (MW), and IR measurements, and limited gauge observations, they inevitably contain errors caused by retrieval algorithms, the measurement technologies, and calibration processes (Yong et al., 2016;Gebregiorgis et al., 2018). Supplementary reanalysis datasets are developed by combining remote sensing products, climate model outputs, and observed data (Sun et al., 2018). Nowadays, many reanalysis datasets are available, such as the ECMWF Interim Re-Analysis (ERA-Interim), the NASA Modern-Era Retrospective Analysis for Research and Applications (MERRA), and the NCEP Climate Forecast System Reanalysis (CFSR), and the number of such reanalysis datasets is still increasing (Saha et al., 2010;Dee et al., 2011;Gelaro et al., 2017). It is worthwhile mentioning that the China Meteorological Assimilation Driving Datasets for the Soil and Water Assessment Tool (SWAT) model (CMADS) over East Asia was recently developed by using the Space-Time Multiscale Analysis System (STMAS), large nested loop data, resampling, and bilinear interpolation methods (Meng and Wang, 2017;Meng et al., 2018). The CMADS precipitation data used CMORPH as the background and merged it with rain gauge data from over 30,000 automatic weather stations located in China. Several studies demonstrated that the latest CMADS precipitation product performs satisfactorily compared to in situ measurements, and can be an alternative option for obtaining precipitation information (Gao et al., 2018;Guo et al., 2018;Liu et al., 2018;Vu et al., 2018;Zhao et al., 2018;Li et al., 2019;Wang et al., 2020). For instance, Gao et al. (2018) evaluated the hydrological application of CMADS against TMPA 3B42V7, NCEP-CFSR, and PERSIANN-CDR over Xiang River basin in China, and concluded that the CMADS and TMPA 3B42V7 simulated streamflow well. Although CMADS precipitation data have been widely evaluated, the performances are inconsistent in different regions. Meanwhile, as CMADS precipitation record is based on the CMORPH SPP, it should be comprehensively compared to TRMM/Global Precipitation Measurement (GPM)-era commonly used SPPs (particularly the CMORPH-BLD product) to analyze the superiority of CMADS for merging rain gauge data.
In this study, five open-access precipitation products, including the newly released reanalysis precipitation dataset (CMADS) and four TRMM/GPM-era commonly used bias-adjusted SPPs (TMPA 3B42V7, CMORPH-BLD, CHIRPS, and PERSIANN-CDR), were assessed. We aim to evaluate the applicability of the five precipitation products over upper Huaihe River basin in China during 2008-2015, and then further analyze the error characteristics of CMADS over the Chinese mainland. The findings can serve as a reference for selecting suitable open-access precipitation datasets for hydrological applications and help to improve future versions of the CMADS precipitation product.

Study area
The Xixian Basin, located in the southwestern part of the Huaihe River basin, extending from 31.5°N to 32.75°N and from 113.25°E to 115°E, has an area of 10,191 km 2 above the Xixian Hydrological Station (Fig.  1). Western and southern edge of the basin are mountainous areas, and low depressed landforms are in downstream, with elevations ranging from 33 to 1110 m. Cropland (41.8%), woodland (38.8%), paddy field (17.2%), grassland (0.54%), and water (1.32%) are the main types of land-use in the basin (Shi et al., 2011). The basin is experiencing average annual precipitation and runoff of 1145 and 371 mm, respectively. As it is affected by the East Asian monsoon in the flood season, precipitation occurs mainly between June and September. The frequent heavy rain increases the risk of flooding, thus, the hydrological application of SPPs in this basin should be assessed (Jiang et al., 2018a).

Reanalysis and SPPs
CMADS is a reanalysis dataset that was developed by using assimilation techniques and processing methods (Meng and Wang, 2017;Meng et al., 2018). To meet the demands for various scientific research and guarantee the reanalysis data with high resolution and high quality, CMADS includes temperature, humidity, solar radiation, wind speed, precipitation, and other variables. Using the CMORPH product as the background field, the CMADS precipitation data for China incorporate the hourly merged precipitation product from the China National Meteorological Information Center. CMADS covers the area of 0°-65°N and 60°-160°E. The CMADS precipitation product with spatiotemporal resolutions of 0.25° and 24 h covering the period of 2008-2015 was used in this study.
Launched by NASA and the Japan Aerospace Exploration Agency in 1997, the TRMM was designed to monitor tropical rainfall. The TMPA, one of the products of the TRMM, was generated from combining passive MW and IR observations (Huffman et al., 2007). The postreal-time TMPA 3B42V7 product, which is available for 50°S-50°N, covers the period of 1998-2015, and provides of 0.25° and 3 h for spatial and temporal resolution. By incorporating the Global Precipitation Climatology Center (GPCC) monthly precipitation data through inverse-error-variance weighting method for bias adjustment, TMPA 3B42V7 has performed well in many precipitation and hydrological utility research (Yong et al., 2010;Xue et al., 2013;Yuan et al., 2017).
The CMORPH product is estimated by the NOAA CPC morphing technique (Joyce et al., 2004). CMORPH Version 1.0, which is the most recent version, includes three products. Raw CMORPH (CMORPH-RAW), only estimated from satellite precipitation, was generated by both passive MW observations and IR data from low-orbit and multiple geostationary satellites, respectively. Bias-corrected CMORPH (CMORPH-CRT) is developed based on the CMORPH-RAW product but using the probability density function (PDF) to adjust the bias. The gauge-SPP, CMORPH-BLD, was outputted based on the CMORPH-CRT product merged with CPC unified daily gauges following an optimal interpolation (OI) method. All three CMORPH products that begin in 1998 are available within 60°S-60°N. Here, the CMORPH-BLD with spatiotemporal resolutions of 0.25° and 24 h, providing data from 2008 to 2015, was employed and evaluated.
CHIRPS is a merged satellite-gauge product that uses the following sources: Climate Hazards Group Precipitation Climatology (CHPClim), geostationary thermal IR satellite observations from both the CPC IR and National Climatic Data Center (NCDC) B1 IR, the TRMM 3B42 product, the rainfall fields of the atmospheric model from the NOAA Climate Forecast System Version 2 (CFSv2), and the gauge observations from multiple sources (Funk et al., 2015). Released in early 2014, CHIRPS has a relatively high spatial resolution of 0.05°, covers the 50°S-50°N region, and provides rainfall data from 1981 to present.
PERSIANN-CDR provides precipitation data with spatiotemporal resolutions of 0.25° and 24 h from 1983 to present, and covers the area between 60°S and 60°N (Sorooshian et al., 2000). An artificial neural network (ANN) model was used to convert IR observations from geostationary satellite imagery into precipitation rates. The Stage-IV precipitation data from the NCEP were used for the initial training of the ANN model, and its parameters remained fixed during the running period. The GridSat-B1 data were used as input for precipitation, producing PERSIANN B1, which was further calibrated by using the monthly Global Precipitation Climatology Project (GPCP) data to reduce bias (Table 1).

Gauge precipitation data
Daily precipitation data were obtained from 22 densely distributed rain gauge stations in the Xixian Basin (Fig. 1). The daily streamflow data for the Xixian Hydrological Station and meteorological data for the Xinyang Station were both collected from the Hydrological Bureau of the Ministry of Water Resources of China. The rain gauge data were converted into spatially distributed precipitation data following the inverse distance weighting interpolation method (Bartier and Keller, 1996). The digital elevation model (DEM) data with a spatial resolution of 30 arc s used in this study were obtained from the U.S. Geological Survey. The vegetationtype data were collected from the Moderate Resolution Imaging Spectroradiometer land cover data using the International Geosphere-Biosphere Program classification system (Friedl et al., 2002).

The Xin'anjiang (XAJ) model
Developed by Zhao (1992) in late 1970s, the XAJ model is a renowned hydrological model and has been broadly applied to simulate streamflow, evaluate water resources, and design hydrological station networks in humid and semiarid watersheds throughout China (Zhao, 1992;Xu et al., 2016;Jiang et al., 2018b). A grid-based XAJ model was used in this study, which uses the saturation excess runoff scheme to compute total runoff of every single grid, and then divides the streamflow into surface runoff, interflow, and groundwater flow. The overland flow concentration is calculated by the flow direction within the grid, and the river network flow concentration is then computed to simulate streamflow. The model contains 16 parameters and their numeric ranges and default values are shown in Section 4.3. Several model parameters are very sensitive and need to be calibrated (Zhao, 1992;Yuan et al., 2019). For instance, the  Fig. 1. Locations of the rain gauges, hydrological station, and meteorological station in the study area, and its topography and geographical position.
OCTOBER 2020 coefficient of potential evapotranspiration (Kc) predominantly controls the simulated total runoff. The areal mean free water storage capacity (SM) represents the depth of the gravitational water reservoir and largely regulates the magnitude of high flow. The outflow coefficients of the free water storage to interflow and groundwater (KI and KG) determine the flow rate from free water storage and the proportion going to interflow and groundwater flow. The recession constants of surface, interflow, and groundwater runoffs (CS, CI, and CG) regulate the recession rate of each runoff component. For the detailed description of the grid-based XAJ model parameters, please refer to Zhao (1992). In this study, the ratio of impervious area (IM) was calculated by the land cover data, and the rest parameters were automatically calibrated by using the shuffled complex evolution algorithm (SCE-UA; Duan et al., 1994).

Statistical indices
The statistical metrics are separated into two groups in terms of their function (Jiang et al., 2018c). The first group includes the Pearson correlation coefficient (CC), root-mean-square error (RMSE), relative bias (BIAS), and mean error (ME). CC describes the linear correlation in satellite/reanalysis datasets and observed data. BI-AS shows systematic bias. RMSE denotes the average error magnitude of the datasets. ME indicates the average differences between the satellite/reanalysis rainfall estim-ates and rainfall observations. The second group includes the probability of detection (POD) and false alarm ratio (FAR), which present the correspondence between the two fields, as well as the capability for detecting precipitation. POD reflects the probability of rain events correctly detected. FAR indicates false cases when satellite/reanalysis records rainfall while no rain happens. Additionally, to evaluate the capability of simulated streamflow that derived by precipitation products, CC, BIAS, and Nash-Sutcliffe efficiency (NSE) were employed. NSE assesses the goodness of the hydrological model based on the correspondence between simulated runoff and ground observations (Duan et al., 2019;Jiang et al., 2019). Table 2 lists the formulae for metrics mentioned above.

Statistical evaluation and comparison of precipitation products
We evaluate the reanalysis and satellite-based precipitation products against rain gauge precipitation from 2008 to 2015. To ensure a comprehensive comparison, the grid pixels precipitation (22 grids) and basin average precipitation are taken to calculate the statistical metrics, respectively. Figure 2 shows the density-colored scatterplots of daily precipitation at the grid scale, which compares the are the evaluated and observed values; and are the mean values of and , respectively; is the number of samples is the number of observed and detected rainfall events; is the number of detected but not observed rainfall events; is the number of observed but not detected rainfall events 1 -  and basin scales are summarized in Table 3. The performance of CMORPH-BLD is better than that of the other four products, as it has high CC and low RMSE values (0.85 and 4.91 mm respectively). The performance of CMADS is comparable to that of CMORPH-BLD, for the CC and RMSE values are slightly better, but with a larger BIAS of −20.5%, indicating that it underestimated precipitation. The finding of CMADS underestimation of precipitation is consistent with some previous studies. For instance, Gao et al. (2018) found average bias of −28.7% for CMADS over Xiang River basin and Zhou et al. (2019) conducted average bias of −12.2% for CMADS over Jinhua River basin, which may be caused by the underestimation of the background field CMORPH data. The performance of TMPA 3B42V7 is slightly poorer than that of CMADS, but it outperforms CHIRPS and PERSIANN-CDR. Among these products, PERSIANN-CDR performs the poorest as it exhibits the lowest CC of 0.37 and highest RMSE of 10.9 mm. Based on the systematic error, all products, excluding CMADS, overestimate precipitation. CMORPH-BLD is the best product for detecting rainfall, with the POD reaching 0.95 and low FAR value of 0.3, followed by CMADS, with a POD of 0.91 and FAR of 0.39. As shown in Fig.  3, CMORPH-BLD and CMADS have similar good performance in three metrics: the CC and POD values for these two products are over 0.8 and 0.9 at all station sites respectively, and the FAR values are below 0.5. These results are obviously better than other products, indicating the high accuracy and reliable performances of CMORPH-BLD and CMADS products, which are consistent with some previous research (Sun et al., 2016;Gao et al., 2018). The better performance of CMORPH-BLD and CMADS than TMPA 3B42V7 may be due to the PDF-OI gauge adjustment procedure to correct satellite precipitation, which results in higher quality and more stable performance (Sun et al., 2016). In contrast, TMPA 3B42V7 adopted a simple gauge adjustment algorithm and exhibits relatively poor performance.
At the basin scale, the performance of all five products remarkably improved from their grid-scale performance, as indicated by the increase in the CC values and decrease in the RMSE values. This agrees with previous findings in different regions that increasing the area scale improved the performance of products (Duan et al., 2016;Jiang et al., 2018c;Maggioni and Massari, 2018). As was the case for the grid-scale evaluation, CMORPH-BLD shows outstanding performance with the best CC and RMSE values of 0.95 and 2.58 mm, respectively, then followed by CMADS, and PERSIANN-CDR performs the worst. CMORPH-BLD is the best product for detecting rainfall, with the POD reaching 0.93 and low FAR value of 0.37, followed by CMADS, with a POD of 0.97 and FAR of 0.48. Therefore, both CMORPH-BLD and CMADS could suitably detect rainfall events. Figure 4 presents the daily precipitation intensities during 2008-2015. The study area predominantly endures light rain (< 1 mm day −1 ), which accounted for a large number of all rainfall events; this is also shown in Fig. 2. The TMPA 3B42V7 product agrees best with the gauge data for most precipitation classes, particularly for non-rainy and light rainfall (1-10 mm) days. CMORPH-BLD and CMADS generally perform similarly; however, CMORPH-BLD tends to overestimate precipitation ranging from 10 to 50 mm day −1 , while CMADS tends to underestimate such rainfall. The underestimation of precipitation over 10 mm day −1 by CMADS leads to a significant negative BIAS for the rainfall estimation (Table 3). CHIRPS tends to underestimate precipitation below 25 mm day −1 , but overestimate heavy and torrential rainfall (> 25 mm day −1 ). The result of PERSIANN-CDR is opposite to that of CHIRPS. Excluding CHIRPS, all other four products tend to underestimate torrential rainfall. Notably, although high-intensity precipitation events account for a small percentage of total, they significantly contribute to the total amount of rainfall. As the runoff generation and separation processes within the hydrological models are highly sensitive to the frequency distribution of precipitation, close frequencies to those of the gauge data can ensure accurate hydrological simulation (Tian et al., 2010;Li et al., 2013;Wang et al., 2017).
Taylor diagrams are used to evaluate the performance of the five products at two scales, and simultaneously show the average standard deviation (SD), CC, and RMSE values (Fig. 5). The diagrams provide a concise statistical summary of the indices, in which the products closer to the point representing the gauge observations perform better than the others. CMADS outperforms other products at both the grid and basin scales. The performance of CMORPH-BLD is comparable, though the SD value is slightly higher, and TMPA 3B42V7 exhibits higher SD and RMSE values. The points representing CHIRPS and PERSIANN-CDR are both located far away from the gauge point, indicating relatively poor performance. As expected, all five precipitation products perform better as the scale increased from the grid to the basin, the same as the results of metric evaluation in Table 3.

Hydrological utility evaluation
The suitability of the five precipitation products for hydrological assessments is evaluated by simulating streamflow. The XAJ model is used under two scenarios.
In Scenario I, model parameters are calibrated by using gauge observations as forcing data for calibration period (2008-2012), and the model is verified during the validation period (2013)(2014)(2015). We then run the model with the five satellite datasets as forcing data by using parameters above.
In Scenario II, the model runs with the same calibration and validation period as Scenario I, except for using the five precipitation data as input data in calibration period, and then individually using the parameters for simulation in validation period.

Daily streamflow simulations under Scenario I
Rain gauge data simulations during the whole periods are shown in Fig. 6. The streamflow simulated by using gauge data agrees well with the observations: the NSE values in calibration and validation periods are 0.86 and 0.78, CC values are 0.93 and 0.89, and BIAS values are 1.31% and 11.0%, respectively. Thus, the XAJ model can be reliably used in the Xixian Basin.
The five precipitation products are then used as forcing data for simulating streamflow (Fig. 7 and Table 4). OCTOBER 2020 Generally, CMADS simulates streamflow most accurately throughout both periods, as it exhibited the highest NSE of 0.81 and 0.69, and highest CC of 0.92 and 0.85 for the calibration and validation periods, respectively. However, the simulations significantly underestimate streamflow during the calibration period, with BIAS of −24.8%, which could be due to the underestimation of heavy and torrential rain. The consistent high bias of both the precipitation estimates and streamflow simulation indicates that the error in the rainfall estimates propagates into the hydrological simulations through the Massari, 2018). For instance, PERSIANN-CDR has significant overestimation (nearly 20%) in light rain (1-10 mm day −1 ; Fig. 4), and therefore leads to positive bias in rainfall. However, the overestimation in light rain and underestimation in heavy rain of PERSIANN-CDR could affect the streamflow simulation while using parameters calibrated by gauge observations, and might lead to underestimation in streamflow.

Daily streamflow simulations under Scenario II
To further evaluate the capability for hydrological simulation, the XAJ model parameters are individually recalibrated by using the five satellite products as input, and the performances are summarized in Fig. 8 and Table 4. All simulations under Scenario II, with higher NSE and CC values, perform better than those under Scenario I, which coincides with previous findings (Jiang et al., 2018c;Maggioni and Massari, 2018). Similar to Scenario I, CMADS outperforms the other products for the entire period, with the highest NSE and CC values: for calibration period they are 0.85 and 0.92, and for validation period they are 0.75 and 0.85, respectively. Unlike the high BIAS under Scenario I, the BIAS exhibits reasonable values under Scenario II, with values of −2.71% in calibration period and 5.10% in validation period, respectively, indicating that the recalibration of model parameters compensates for the rainfall underestimation by the products, thus improving the hydrological performance of the model. The CMADS-calibrated model captures most of the peak flows, as shown in the scatterplots in Fig. 8. However, the simulations tend to underestimate discharge when capturing high peak flows, which is partly due to the uncertainty in the detection of heavy and torrential rainfall events by CMADS as shown in Fig. 4.

Effects of model recalibration on streamflow simulation
The effects of recalibration on model parameters and streamflow simulation when using specific precipitation data as input are further analyzed. Table 5 shows the parameters of the grid-based XAJ model calibrated with gauge observation and each specific precipitation data as input for the calibration period 2008-2012. The Kc is a sensitive parameter that can predominantly control the simulated total runoff. When the Kc value is increased, the calculated evapotranspiration also increases; conversely, the evapotranspiration decreases. For the significant negative bias of the streamflow simulations for the CMADS, CMORPH-BLD, and PERSIANN-CDR products, the recalibration reduced the Kc values from 1.49 to 0.96, 1.33, and 1.25 to achieve a good simulation of the streamflow (see Table 4), respectively. The SM predominantly regulates the high flow magnitude, and reducing SM tends to partition a larger proportion of surface runoff from the total runoff and hence produces high flood peaks (Zhao, 1992). Under Scenario II, the model recalibration alleviates the magnitudes of high-flow underestimation (Figs. 8a, b,   OCTOBER 2020 portion going to interflow and groundwater flow, and a larger KI can produce higher flood peaks. The model recalibration alleviates the magnitudes of high-flow underestimation (Fig. 8) via increasing KI from 0.36 to 0.37 for CMADS, 0.41 for CMORPH-BLD, and 0.44 for PERSIANN-CDR. Moreover, reducing the recession constants of surface, interflow, and groundwater runoffs (CS, CI, and CG) may lead to a faster recession rate for each runoff component and consequently augment high flows to a certain extent. Table 5 shows that the CS value of CMORPH-BLD and CI values of five precipitation products had reduced, thereby partially compensating for the high-flow underestimation. By recalibration, the model sensitive parameters are mostly changed into the direction that makes the simulation better. However, it should be noted that for the complexity and uncertainty of the model parameters, not all the recalibrated parameters of specific precipitation product changed consistently, which could be attributed to the equifinality for different parameters (Beven and Freer, 2001). Model recalibration has also been proved to be a viable strategy for improving satellite and reanalysis precipitation-forced hydrological model performances in some other basins of different size and climatology in Ethiopia, China, and Myanmar (Bitew and Gebremichael, 2011;Jiang et al., 2012;Maggioni and Massari, 2018;Yuan et al., 2019). The reason can be attributed to the different precision characteristics of the satellite and reanalysis precipitation with respect to the gauge measurements. The recalibrated parameter settings can compensate for the inaccurate satellite and reanalysis precipitation errors to some extent, thereby improving streamflow simulations with respect to the gauge observations calibration option (see Tables 4 and 5). However, it is worth noting that the recalibration cannot completely compensate for the errors from inaccurate precipitation forcing input. After the parameter recalibration, the most inaccurate precipitation (i.e., the PERSIANN-CDR precipitation product in this study) still had the largest streamflow simulation error in terms of CC and NSE values (see Table 4). This could be related to the threshold gain function in error transfer of the hydrological model and parameter recalibration; that is, if the precipitation input error is higher than a certain threshold, the streamflow output error will be hard to eliminate, and if the precipitation input error is lower than a certain threshold, the runoff output error will not increase much, or even decline (Yong et al., 2010;Jiang et al., 2012;Mei et al., 2016).

Comparison with the findings of other CMADS studies
The results of recent studies evaluating the applicability of the CMADS precipitation product to simulate streamflow over the Chinese mainland are summarized in Table 6. For precipitation estimation, based on the CC values, the accuracy of the CMADS precipitation product in humid regions (such as the Xiang and Jinhua River basins, and Huaihe River source basin) is higher than that in semiarid regions (such as the Yellow River source basin), indicating that the CMADS could estimate precipitation in the humid regions of Chinese mainland well. In terms of the BIAS values, the CMADS precipitation product tends to underestimate precipitation in humid regions, but overestimate precipitation in semiarid regions. The bias in the background data and assimilation algorithm, unevenly distributed rain gauge stations, and the complex topography of some regions may have attributed to the high bias of the CMADS precipitation estimates. For streamflow simulation, most studies (except for Jinhua River and Xihe River basins) reported good performance with satisfactory NSE and BIAS values owing to parameter recalibration, which compensated for the bias in the rainfall results. Overall, these results demonstrate the outstanding hydrological performance of the CMADS precipitation product, and suggest that CMADS is suitable for various regions in China, particularly humid regions.

Performance of CMADS precipitation estimates
for the Chinese mainland Owing to the remarkable performance of CMADS for estimating precipitation and simulating streamflow in recent studies, its applicability in different climates and elevation bands should be assessed. We further evaluate its performance for the Chinese mainland by calculating the CC, BIAS, POD, and FAR against data from 824 rain gauges collected during 2008-2015. Figure 9 shows the rain gauge distribution which is relatively evenly spaced throughout the Chinese mainland. Figure 9 plots the spatial distributions of the metrics for the CMADS estimates over Chinese mainland. The distributions of CC and POD are identical, and declined from southeast to northwest. This trend is also similar to the precipitation intensity distribution over the Chinese mainland, i.e., the decline of CC and POD values coincides with the declination of the precipitation intensity. The CC and POD of CMADS are generally high in eastern and southwestern China, where heavy rainfall events frequently occur. However, the CC and POD values are lower in northwestern China. The complexity of the climate and topography in such areas challenges the accur-acy of satellite precipitation estimates, particularly in mountainous regions (i.e., the Tibetan Plateau), which resulted in the low CC and POD values. Moreover, interpolated rainfall errors are more likely to occur in areas with a sparse gauge network, increasing the uncertainty of precipitation estimates. Therefore, apart from the satellite retrieval algorithm, the low accuracy of precipitation estimates over these areas could also be due to the sparse gauge network. The FAR demonstrates similar spatial distribution of precipitation precision with CC and POD. The high POD and low FAR values in southeastern China indicate that CMADS could detect and capture rainfall events in this region. Similar to our study area, CMADS underestimates the precipitation over most regions with relatively high bias, as shown in Fig. 9b. CMADS tends to underestimate precipitation across most regions of China, and overestimate precipitation in areas with limited rainfall and sparsely distributed rain gauge stations. Owing to the significant underestimation of rainfall in most regions of China, researchers should proceed with caution when applying CMADS precipitation product to simulate and predict hydrological extremes.

Conclusions
In this study, the performance of one recently released reanalysis precipitation dataset (CMADS) and four biasadjusted satellite precipitation datasets (TMPA 3B42V7, CMORPH-BLD, CHIRPS, and PERSIANN-CDR) was evaluated and compared against rain gauge observations at the grid and basin scales. The performance of these products in driving the XAJ model for streamflow simulation was also assessed. The major conclusions are summarized as follows.
For statistical assessment, the CMADS reanalysis and CMORPH-BLD satellite precipitation datasets both perform similarly, with high CC and low RMSE values. Notably, CMADS generally underestimates precipitation while the other products present reasonable bias. Generally, CMADS and CMORPH-BLD both perform the best, followed by TMPA 3B42V7, while CHIRPS and PERSIANN-CDR perform poorly. CMORPH-BLD performs the best in capturing and detecting rainfall events. While CMADS tends to underestimate heavy and torren-  Zhao (1992) and Yuan et al. (2019). tial precipitation, and the underestimation may have contributed to the significant bias in the evaluation of precipitation.
The results of the streamflow simulations under Scenario I show that CMADS outperforms CMORPH-BLD. Under Scenario II, the simulation performances of all five precipitation products are significantly better than those under Scenario I. Despite the high negative bias in the statistical evaluation of CMADS precipitation, the bias of streamflow simulation under Scenario II is reasonable, demonstrating that the CMADS precipitation product is applicable to hydrological evaluation and simulation.
The spatial distribution of the evaluation metrics of CMADS over Chinese mainland is similar to that of precipitation intensity, which exhibits high accuracy in eastern China and declines from southeast to northwest. Despite the rain gauge density, performance of the metrics is also affected by the climate and elevation conditions, resulting in the poor performance in northwestern China.
Overall, the CMADS reanalysis precipitation dataset could provide reasonably good rainfall estimation and can drive hydrological model to generate good runoff simulation in the Xixian Basin. CMORPH-BLD and CMADS products show similar performance and are comparable with each other, as the CMADS precipitation product uses CMORPH as a background field. However, owing to the high negative BIAS values of CMADS precipitation data, the CMADS research community should further improve the calibration algorithms and enhance the quality of the precipitation product for eastern Asia. In addition, this study only evaluated CMADS against a limited number of SPPs. In the future, we should further strengthen the scientific evaluation and practical application of CMADS, such as comparing CMADS with the latest Integrated Multi-satellite Retrievals for GPM Version 6 products (Huffman et al., 2019).